AuroraASR-1M: An Extremely Lightweight 1MB Speech Recognition Model
AuroraASR-1M has been released as an open-source, lightweight automatic speech recognition model containing 949K parameters and a file size of just 1.06 MB for INT8 weights. It utilizes a weight-shared Conformer-CTC architecture to achieve fast inference speeds on standard CPUs. This release provides a highly efficient tool for edge devices and local AI applications where memory and computational resources are severely constrained. By running at approximately 436 times real-time on a single CPU thread, it opens up new possibilities for embedded and offline speech recognition. The model achieves an 11.43% Word Error Rate (WER) on the LibriSpeech test-clean dataset using greedy decoding without an external language model. Furthermore, it is distributed under the permissive Apache 2.0 license.
## BACKGROUND
Automatic speech recognition (ASR) models translate spoken language into text, often evaluated using Word Error Rate (WER) where lower percentages indicate better accuracy. Architectures like Conformer-CTC combine convolutional neural networks and transformers to process speech efficiently, while INT8 quantization reduces memory usage by lowering numerical precision.